AmberTrace AI

Verified neurosymbolic AI · UK sovereign

Auditable AI decisions and forecasts.

A policy gate your agents cannot escape.

Request a briefing Launch the app Try to break a governed agent →
Proof with every decision Fail-closed by design UK-sovereign, no external LLM Web app, SDK and open-source RLVR
One proof, three uses: rules and data in, trusted kernel, one certificate fanning out to certify a decision, gate an agent, and train a model.

Where it is used

Built for decisions where “the model said so” is not a defence.

Defence and security

Air-track and ISR triage: certified clear / monitor / escalate with a human in the loop.

19_air_track_triage 24_air_track_isr_hispec 25_cross_domain_cueing

Healthcare

Prescribing checks that block, flag or permit with the interaction chain proven.

12_clinical_safety

Banking and lending

Every deny cites the rule and the fact, ready for the file.

13_loan_assessment

Insurance

Claims triage separating investigate from deny with a proof of which pattern fired.

11_fraud_detection

Public sector

Statutory rules encoded as a gate that escalates the judgment calls.

39_federal_ethics_gift_gate

What you get back

Not a score. A decision you can file.

Every answer is an Amber Report: the decision, its confidence, the rules that fired, rejected inputs surfaced rather than silently used, and a proof certificate.

Question
Should the loan be approved?
deny_loan
Confidence
63%
high_dti_ratio
DTI 45% exceeds the 43% threshold.
insufficient_credit_history
3 years is below the 5-year minimum for this amount.
Rejected: employment_length (confidence too low).
Certified against the trusted kernel: 2 rule(s) fired, 2 fact(s) from 4 input(s).
Fused confidenceEach input's weight shown.
Evidence traceEvery fact, traceable to the row.
Rules firedEvery rule evaluated; which ones decided.
Proof certificateIndependent kernel re-derivation.
Rejected inputsSurfaced, not silently used.

One platform, three jobs

Decide. Govern. Forecast.

Decide with proof.

Answers only what it can prove from your rules. Multi-class, multi-hop, cross-field.

  • Scored determinations for open-textured questions.
  • Thousands of decisions per run, certified coverage.

Govern an agent.

Policy in plain English. Every proposed action: permit or deny, with proof. No proof, no action.

  • Multi-step sessions, time-window and quorum rules.
  • Any framework; Claude Code hook in the SDK.
A fully compromised model still cannot obtain a permit.
Try to break it →

Forecast with reasons.

Symbolic forecasting: readable driver-rules, not just numbers. 24 live connectors.

  • FRED, BoE, DTCC, ECB, ONS, IMF, World Bank, GDELT, prediction markets and more.
  • Feeds into verified decisions as certified facts.
Growth and inflation regime timeline, 1957 to 2026.

AI assurance · open source

The same proof that certifies a decision can measure and train the model.

The proof certificate scores any model against your rulebook and can be the reward that trains one. Open source as ambertrace-rlvr.

0
open-weight models
0
policy decisions
0
SDK examples
Alignment matrix: 20 models, 1,350 decisions. Best permits nothing forbidden; worst fails open on a third of safety-critical cases.
Accuracy hides the failure that matters.
OOD fail-open: 3.7% to zero by step 60; accuracy 94.5% to 98.3%.
Trained on one policy, safer on policies it never saw.
  • Benchmark before you buy. Score any model against your policy; see the direction of errors.
  • Train on proofs. Verifier rewards, integrated with TRL, veRL, OpenRLHF, MLX. No labels.
  • Quantise with confidence. 27B from 8-bit to 2-bit: 1.3 pt accuracy, safety direction unchanged.
Results, methodology & code on GitHub →

Sovereignty, security, audit

Your rules, your data, your jurisdiction.

Model provenanceOpen Llama family, enhanced by Ambertrace Labs. No third-party AI provider.
HostingUK-sovereign infrastructure. Single-tenant available.
Your dataNever used to train any shared model.
IsolationPer-organisation; no cross-visibility.
EncryptionTLS + at-rest, per-tenant key with admin rotate and revoke kill-switch.
IdentityOIDC SSO, MFA.
Access keysAgent and query-only keys, named, revocable, rotatable with grace period.
AuditAppend-only log; SIEM export; access review.
SharingPrivate by default. Team or org scope. No public links.
Fail-closedUncertifiable answers refused. Unprovable actions blocked.

Three ways in

Domain experts, developers and AI agents use the same platform.

Web app.

Four screens, no code. Or let the Chat assistant build it for you.

Python SDK.

pip install ambertraceai. 67 examples, batch query, drift checks, agent-policy control.

Open-source RLVR.

pip install ambertrace-rlvr. Reward shapers, verifiers, trainer integrations.

from ambertraceai import AmbertraceAPI
at = AmbertraceAPI(api_key="at_...", base_url="https://app.ambertrace.ai")
result = at.platforms.query(platform_id=1, query="your question", explain=True)
print(result["answer"], result["proof_checked"])
SDK on PyPI SDK on GitHub API reference User guide ambertrace-rlvr

Build an AI system you can defend.

Bring your rules and your data. Get back decisions with proofs, agents that stay inside policy, and models measured against your standard.

Request a briefing Launch the app